Attribute preference and priming in reference production: Experimental evidence and computational modeling - eScholarship
Albert Gatt, Martjin Goudbeek, Emiel Krahmer · Proceedings of the Annual Meeting of the Cognitive Science Society · 2011
Attribute preference and priming in reference production Experimental evidence and computational modeling Albert Gatt ([email protected]) Institute of Linguistics, University of Malta Tilburg Center for Cognition and Communication (TiCC), Tilburg University Martijn Goudbeek ([email protected]) Tilburg Center for Cognition and Communication (TiCC), Tilburg University Emiel Krahmer ([email protected]) Tilburg Center for Cognition and Communication (TiCC), Tilburg University Abstract Referring expressions (such as the red chair facing right) of- ten show evidence of preferences (Pechmann, 1989; Belke & Meyer, 2002), with some attributes (e.g. colour) being more frequent and more often included when they are not required, leading to overspecified references. This observation underlies many computational models of Referring Expression Genera- tion, especially those influenced by Dale & Reiter’s (1995) In- cremental Algorithm. However, more recent work has shown that in interactive settings, priming can alter preferences. This paper provides further experimental evidence for these phe- nomena, and proposes a new computational model that in- corporates both attribute preferences and priming effects. We show that the model provides an excellent match to human ex- perimental data. Keywords: Reference, production, Natural Language Gener- ation, Computational Modeling Figure 1: A referential domain Introduction In domains such as Figure 1, where a target referent needs to be distinguished from its distractors in context, people of- ten produce overspecified descriptions such as the red sofa facing right, when a description containing fewer attributes would suffice (Pechmann, 1989; Eikmeyer & Ahls`en, 1996; Belke & Meyer, 2002; Engelhardt, Bailey, & Ferreira, 2006). This finding challenges the assumption that speakers observe the Gricean Maxim of Quantity by not including any more in- formation than is relevant for identification (cf. Olson, 1970, for an early adoption of this view). One important observation in this regard is that certain at- tributes (for example, an object’s colour), are more likely to be redundantly included in an overspecified description than others (such as size or orientation) (Pechmann, 1989; Belke & Meyer, 2002). The preferred status of such attributes may arise due to their perceptual salience, higher codability rel- ative to other attributes (Belke & Meyer, 2002) and/or be- cause they form an integral part of the conceptual represen- tation of the object (Pechmann, 1989). On one interpretation of these findings, preferred attributes are selected first when a description is being formulated; since this is an incremen- tal process, should later attributes be included which make them redundant, the whole description would be overspeci- fied (Pechmann, 1989; Levelt, 1989). This has important implications for computational mod- els of referring expressions generation ( REG ), which seek to model the process of attribute selection for identifying descriptions. Such models form an integral part of Nat- ural Language Generation systems, which generate text or speech from non-linguistic input. Current REG models per- form attribute selection primarily on the basis of discrimi- natory value: does a target attribute help to exclude some distractors in the domain? Some models (e.g. Dale, 1989; Gardent, 2002) seek to satisfy a strict interpretation of the Gricean maxim of quantity by selecting the smallest set of at- tributes that would uniquely identify the target referent(s). An alternative, more influential model is Dale and Reiter’s (1995) Incremental Algorithm, which is in part inspired by the psy- cholinguistic literature and models attribute selection as an incremental search that prioritises more preferred attributes. As we show below, such models can overspecify in some sit- uations. Furthermore, they have been shown to match speaker behaviour better than earlier models (Gatt, van der Sluis, & van Deemter, 2007; Gatt & Belz, 2010). Many of the psycholinguistic studies cited above were undertaken in non-interactive settings, whereas recent psy- cholinguistic work on dialogue has highlighted the extent to which speakers’s production choices are influenced by their interlocutors’. One aspect of this process, discussed by Clark et al. (Clark & Wilkes-Gibbs, 1986; Brennan. & Clark, 1996), is the ‘negotiation’ on the best way to refer to an ob- ject that characterises some interactive reference tasks. More recently, Pickering and Garrod (2004) have proposed the In- teractive Alignment model, whereby interlocutors ‘align’ at various levels (for example, syntactic and semantic) as a re- sult of a basic priming mechanism. There is substantial evi- dence that such priming occurs, particularly in interlocutors’